1. Field of the Invention
The present invention relates, in general, to distributed computing and clustered computing environments, and, more particularly, to computer software, hardware, and computer-based methods for hosting a set of computer clusters that are uniquely configured or customized to suit a number of remote customers or clients.
2. Relevant Background.
A growing trend in the field of distributed computing is to use two or more computing resources to perform computing tasks. These grouped resources are often labeled clustered computing environments or computing clusters or simply “clusters.” A cluster may include a computer or processors, network or communication links for transferring data among the grouped resources, data storage, and other devices to perform one or more assigned computing processes or tasks. The clusters may be configured for high availability, for higher performance, or to suit other functional parameters. In a typical arrangement, a portion of a company's data center may be arranged and configured to operate as a cluster to perform one task or support the needs of a division or portion of the company. While a company may benefit from use of a cluster periodically on an ongoing basis, there are a number of reasons why it is often undesirable for a company to own and maintain a cluster.
As one example, High Performance Computing (HPC) clusters are difficult to setup, configure, and manage. An HPC cluster also requires numerous resources for ongoing maintenance that increases the cost and manpower associated with cluster ownership. Despite these issues, a company may require or at least demand HPC clusters (or other cluster types) to solve large problems that would take an inordinate amount of time to solve with a single computer. The need for HPC and other cluster types is in part due to the fact that processor speeds have stagnated over the past few years. As a result, many companies and other organizations now turn to HPC clusters because their problems cannot be solved more rapidly by simply purchasing a faster processor. These computer users are placed in the difficult position of weighing the benefits of HPC clusters against the resources consumed by owning such clusters. Decision makers often solve this dilemma by not purchasing clusters, and clusters have remained out of reach of some clients as the resource issues appear insurmountable.
When utilized, HPC systems allow a set of computers to work together to solve a single problem. The large problem is broken down into smaller independent tasks that are assigned to individual computers in the cluster allowing the large problem to be solved faster. Assigning the independent tasks to the computer is often the responsibility of a single node in the cluster designated the master node. The responsibilities of the master node include assigning tasks to nodes, keeping track of which nodes are working on which tasks, and consolidating the results from the individual nodes. The master node is also responsible for determining if a node fails and assigning the task of the failed node to another node to ensure that node failures are handled transparently. Communication between nodes is accomplished through a message passing mechanism implemented by every member of the cluster. Message passing allows the individual computers to share information about their status on solving their piece of the problem and return results to the master node. Currently, those who determine a cluster is worth the drain on resources purchase a cluster, host the cluster, and manage it on their premises or on site.
Unfortunately, while the number of tasks and computing situations that would benefit from HPC clusters continues to rapidly grow, HPC clusters are not being widely adopted. In part, this is because HPC clusters require the most computers of any cluster type and, thus, cause the most problems with maintenance and management. Other types of clusters that have been more widely adopted include the “load balancing cluster” and the “high availability cluster,” but resources are also an issue with these clusters. A load balancing cluster is a configuration in which a server sends small individual tasks to a cluster of additional servers when it is overloaded. The high availability cluster is a configuration in which a first server watches a second server and if the second server fails, then the first server takes over the function of the second server.
The multi-cluster subsumes all other classes of clusters because it incorporates multiple clusters to perform tasks. The difficulties for managing clusters are amplified when considering multiple clusters because of their complexity. For example, if one HPC cluster consumes a set of resources, then multiple HPC clusters will, of course, consume a much larger set of resources and be even more expensive to maintain. One method proposed for managing multiple high availability clusters is described in U.S. Pat. No. 6,438,705, but this method is specific only to the managing of high availability clusters. Further, the described method requires each cluster to have a uniform design. Because it is limited to high availability clusters, the owner would not have an option to incorporate multiple cluster types, such as HPC or load-balancing clusters, within the managed multi-cluster. Additionally, the suggested method does not solve one of the fundamental difficulties associated with cluster usage because it requires the cluster to be owned and operated by the user and to remain on the client's property or site. Other discussions of cluster management, such as those found in U.S. Pat. Nos. 6,748,429, 5,371,852, and 5,946,463 generally describe a single cluster configuration and do not relate to operating multi-clusters. In all of these cases, the burden of managing, monitoring, and hosting the cluster remains with the user of the cluster who owns the cluster who must maintain the cluster on their premises.
Hence, there remains a need for systems and methods for providing clusters to users or “clients” such as companies and other organizations that provide the computational assets or power that the clients demand while not presenting an unacceptable burden on the clients' resources. Preferably, these systems and methods would be effective in providing a cluster that is adapted to suit a particular need or computing task rather than forcing a one-size-fits-all solution upon a cluster user.